Executive brief
vLLM is a high-performance engine used for serving large language models. In multi-node configurations, a communication component incorrectly exposes a network port to all interfaces without authentication. An attacker could connect to this port to view internal system metadata or repeatedly connect to cause a denial-of-service, potentially disrupting AI model inference operations.
Technical details
In multi-node vLLM deployments using tensor parallelism, the primary host opens a ZeroMQ XPUB socket and binds it to all available network interfaces (0.0.0.0). This socket is used by the GroupCoordinator to broadcast metadata via the MessageQueue. Any unauthenticated attacker with network access to the host can connect to this socket to receive internal state metadata. Furthermore, by establishing numerous connections and failing to read the published data, an attacker can trigger a denial-of-service condition by blocking the ZeroMQ publisher. The vulnerability is addressed in version 0.8.5.
Affected products
- vLLM Project vLLM >= 0.5.2, < 0.8.5
Timeline
- 2025-04-29: disclosed
- 2025-04-29: advisory
- 2025-04-29: patched